When I Realized I Was Building Something That Fit Nowhere: The Blank Spot on the Category Map

This is the third chapter of a serialized book following one question: in domains that have no right answer, can AI build a kind of learning that makes a person less certain rather than more? I call this project Moral Mirror. The last chapter was about how, when I sat down to give this intuition a form, the first thing I did was subtract. I narrowed to a single trolley, and pared away subscriptions, typing, K-12, and professional ethics. Subtraction makes the outline, I wrote.

But once the outline appeared, a strange thing happened. I could not figure out what to even call the thing left in my hand. Not an education app that teaches the right answer, not a companion app that holds people, not a test that types personality. Whatever box I tried, one side stuck out.

This chapter is a record of finding that blank spot, and at the same time of weighing whether it was an opportunity or just a trap that no one built in for a reason. Let me say up front that the conclusion is not clean. My gut leaned one way, but I could not prove that lean all the way through. That half-settled state is the whole of this chapter.

They Said AI Would Change Everything

To talk about the blank spot, I have to start with why I was looking in education at all. Behind it was a frustration that followed me for a while.

I sometimes think this. AI is a thing that could change, at the root, how people work, how they approach problems, how they think, yet next to that potential it has so far changed rather little. The tools clearly got more powerful, but where that power lands is mostly in doing existing work a little faster, a little cheaper. Write the report faster, write the code faster, draw the picture faster. All good things. But if that is all, then we are bolting a new engine onto a carriage and rolling the carriage a bit quicker. What an engine can change is not the carriage’s speed but the shape of the vehicle.

In education this frustration stands out more. Much of the learning tooling of the past few years shares one skeleton. Summarize a lecture, auto-generate problems, point out where you went wrong. All useful. But that usefulness is on the side of speeding up existing learning, not on the side of changing the shape of learning itself. There are plenty of tools to help you solve multiple-choice faster, but a kind of learning where the multiple-choice format itself has vanished is hard to find. The engine is new, yet it is still pulling a carriage.

Carrying that frustration, my eye went deeper into the domain of learning. Is there another spot where AI could truly make things different, as much as the process by which a person takes something in and changes their own thinking? So I searched. Who is already doing this kind of learning. The kind, in a domain without a right answer, that makes a person see the weak spots in their own thinking for themselves. The more I searched, the clearer it became that it was not in the place it should have been. There were many products that claimed to do something similar, but open them up and they were off, somewhere, from what I had in mind. That gap is where this chapter begins.

I Tried to File It Under Education

At first I naturally thought education. It helps a person learn, so it would slot into edtech (education technology) somewhere. So I tried to put it in that box.

The moment I did, one side stuck out. At the center of what we usually call education, there is a right answer. A math problem has an answer, an English word has a meaning, history has dates. Education is mostly the game of how accurately and efficiently you move that right answer into a head. Learn the concept, practice it, memorize it, confirm it on a test. Good edtech makes this process faster and more personalized. It points out where I got it wrong and drills me on the weak unit. There is a measurable goal, so you can also tell whether the thing was built well.

But what Moral Mirror tries to do is not that game. The trolley has no answer to memorize. Doing well here is not getting more things right but wavering at the spot you believed you were right. Becoming less certain, not more. So it ran against the very center of education’s definition, “move the right answer efficiently.” Even the metric of success was reversed. One side counts getting more right and growing more certain as doing well, the other counts becoming less certain as doing well.

This mismatch spills into measurement too. One reason edtech works well is that you can put a number on whether it worked. Did the accuracy go up, did they move to the next unit, did the test score improve. With that number you can compare plan A and plan B and refine the product toward the better one. But “did they become less certain” is hard to measure that way. By what score do you capture that a person saw the weak spot in their own position? Measure it wrong and it gets dangerous. Take user satisfaction or whether they came back as the metric, and as I wrote in chapter 2, the product slides toward giving comforting lines rather than uncomfortable truths. A box that is easy to measure and a box that is hard to measure are different kinds of thing from the start, and edtech’s toolkit was built for the box that is easy to measure.

The more I researched, the bigger this mismatch grew. Even the learning apps that look similar were, in the end, keeping an answer key somewhere and walking the person toward it. The wording was “personalized learning” and “Socratic tutor,” but open it up and it was a structure that gently steered the learner toward a predetermined conclusion. The education box was the closest box, but the nail at its center, the right answer, was not in mine. With no nail, it would not hang on that box.

But This Is the Same Thing They Did at Harvard

Here I paused. There was one thing that snagged when I tried to say it was not education. What happened in Sandel’s classroom was, when you think about it, the same thing.

Recall what Michael Sandel did on the Harvard campus. He does not make students memorize answers. He throws out the trolley, and when a student answers, he throws the next question that jabs the weak spot in that answer. The student trips over the contradiction inside themselves and leaves the room less certain than they entered. No tests, no answer key. Yet no one calls that “not education.” If anything it is the oldest and deepest form of education, the direct line of the question-and-answer Socrates practiced in the Athenian agora. For thousands of years people have called this learning.

The lineage runs long. The disputatio of the medieval university (a debate format that argued a set thesis back and forth, for and against), the Socratic method of the law school, all descend from this. Learning that shakes a person so they see their own contradiction is not the fringe of education but its deepest mainstream. And yet this never moved into digital products, staying instead in the expensive offline rooms of the lecture hall and the seminar. The saying that to get a good discussion-based class you have to go to a good university means, flipped around, that this learning has never once been unlocked at scale.

So what I was trying to build was not something newly invented. It was education already thousands of years old, with only that box left blank on today’s category map. The boxes for moving the right answer are packed full, while the box for shaking a person had no name in the world of products. Strange. A thing that happens perfectly well in a Harvard classroom has not even a name on the map of apps and services.

This realization changed the nature of the blank spot. It was blank not because there was nothing to put there, but because what should be there was missing. This difference mattered. If it were a strange new field no one had made, the first thing to doubt would be whether it is even needed. But this was a thing that had already proven its value for thousands of years, somehow just never carried over into product. The question of need was already answered, and what remained was “why was it never carried over.” Following that question, the identity of the frustration came into view.

It Is Not That Edtech Was Lazy

So why was it never carried over. At first I nearly thought edtech took the easy road, picking only the domains with right answers. But that was not a fair diagnosis. A diagnosis that calls others lazy usually hides a constraint I failed to see.

To move a Sandel-style classroom into a product, one thing is required. Generating, on the spot, the next question fitted to what the student just said. The thing I wrote about at length in chapter 1. But until recently this was technically impossible. Generating scenarios differently for each person, fresh each moment, required attaching a human facilitator one-on-one, and that did not scale. One Sandel cannot face every learner in the world one by one.

So for someone building an education product, the options were effectively one. Organize pre-made content well, group learners into a few types, and have them follow a normalized path. Even what was called personalization was, in fact, choosing one among pre-built branches. This was not laziness but constraint. And their going to domains with right answers was a natural choice too, because there the normalization worked. Math is taught fairly well even with a pre-made problem bank and an answer key. They used normalization where normalization worked.

The problem is that in domains without a right answer that approach does not work. If you cannot jab the weak spot that differs from person to person, on the spot, then it is not Sandel’s classroom but reading an ethics textbook aloud. The moment you normalize, the core dies. So that spot was bound to stay empty. What made the blank spot was not anyone’s indifference but the wall that bound everyone alike. The wall, from chapter 1’s talk of segments and VR, that you cannot make something fitted to one person separately for a million people.

And chapter 1’s conclusion snags here again. That wall came down. The unit cost of generating reasoning fitted to one person, fresh each moment, dropped to near zero. So this blank spot was not a spot doomed to stay empty, but a spot that became fillable only as the wall fell. The box was missing from the map not because it was useless, but because the tool to fill it did not yet exist in the world. Now that the tool exists, drawing the box has become possible for the first time.

When I Saw the Blank Spot, I Felt a Pull

By the time I got here, honestly my mind had tilted one way. When I learned that it fit nowhere, that no one was really doing it, what I felt was not anxiety but a pull.

My thought went straight there. Oh, then shall I try it, fast. I want to build this and use it. And one more thing, a somewhat personal feeling. I want to study with this a bit more myself. I wanted to go through the experience of wavering in front of the trolley, properly, before anyone else. The wish to be the maker and at the same time the first user. Honestly this came before any business judgment. Before weighing whether there is a market or whether it makes money, the feeling that I would use this if it existed came first.

People who see a blank spot react in two ways, I think. One is “why is this empty, is it not dangerous,” and the other is “this is empty, shall I go in.” I was the second. This pull itself says something about me. I am drawn less to doing better inside an already-defined box, and more to making a box where there is none. That is my disposition.

But disposition is an asset and a trap at once. Let the pull cloud the judgment and you jump in without seeing the signal that there is a reason it was empty. The cases where someone believed they made a new category and it turned out to be a spot with no market are countless. So apart from the pull, I had to weigh once more, coldly, whether this blank spot was a real opportunity or a trap. Trust the pull, but do not trust the pull alone.

Is the Blank Spot an Opportunity or a Trap

An empty market has two faces. That no one is there can mean there is no competition, but it can also mean no one made money or no one wanted it. The first is an opportunity, the second a trap. From a distance both look like the same empty spot. So the thing to do in front of a blank spot is not to decide whether to enter, but to first sort out why it is empty.

%%{init: {'theme':'neutral', 'look':'handDrawn'}}%%
flowchart TD
  E[Blank spot on the category map] --> Q{Why is it empty}
  Q -->|No one wanted it| T[A trap with no demand]
  Q -->|No one could build it| O[An opportunity once supply-blocked]
  O --> W[Has that wall fallen]
  W -->|Fallen = inference-time generation| Y[Fillable now]
  W -->|Still standing| N[Still empty for now]

I tried to split the reasons for the emptiness like this. Some spots are empty because there is no demand. People do not want it, so it does not sell even if you make it. Some spots are empty because until now it could not be made. The demand is there, but the technology to supply it was not. What I sorted out in the chapter before was exactly the second. The demand for one-on-one learning that shakes a person was clearly there. People have replayed Sandel’s lectures tens of millions of times, argue over the trolley at the dinner table, and wonder whether they are a consistent person. What was missing was not the demand but the way to fill that demand fitted to each person. So this blank spot leaned more toward opportunity than trap, was my tentative conclusion.

I also took the trap warning seriously. The products that believed they made a new category and then vanished are many enough to fill a graveyard. Cases that looked like a fine idea, yet no one would actually stand in that spot. What separates such a trap from a real opportunity was, in the end, “what was the reason for the emptiness.” If it is empty because no one wanted it, no improvement in the tool will fill it. But the signals around the trolley were not on the no-demand side. A single public lecture on justice has been played tens of millions of times worldwide, people quarrel over the trolley with friends, and they are curious about the situations in which they change their stance. The hunger for the experience of wavering was everywhere. What was missing was only the supply to fill that hunger fitted to each person.

Still, there is a gap to leave honestly here. Wanting something and opening your wallet for it are different. Not many pay to become uncomfortable. Between enjoying Sandel’s lectures and paying for the experience of being shaken, there is a river to cross. A lecture watched for free and discomfort paid for are different things. I touched this once in chapter 2 talking about the revenue model, and it is a spot I have not yet answered. So my stance now is this. The weight is on the opportunity side, but I have not erased the possibility of a trap. What turns a blank spot into an opportunity is not the fact that it is empty, but proof that the reason it was empty is now gone. I saw that the wall blocking supply came down, but I have not yet seen demand turn into a wallet. I hold only half the proof.

The Two Faces of Having No Name

Once I had made up my mind to go into the blank spot, the next problem followed at once. Having no name was a double-edged thing in itself.

One edge is this. With no name, it is hard to explain to people. To the question “what is this” you cannot answer in one word. Say education and they picture teaching the right answer, say game and they take it lightly, say personality test and they expect type-sorting. All mismatched expectations. A thing with no name is easily called by the wrong name, and called wrong, it gets used wrong. A user who enters expecting a right answer is let down when there is none, and one who enters expecting a light game is flustered when they find themselves shaken. This is a difficulty of experience design before it is a difficulty of marketing.

The other edge is that this very namelessness can become a moat (a defensive line that keeps competitors from easily crossing over). Whoever gets to decide what a thing is called also gets to set the rules of that spot. Opening a new category is hard, but once it is open and the name is accepted, those who follow move on the map I drew. To occupy the spot in “oh, that thing, like such-and-such” when people speak. That is an edge that a feature or two cannot catch up to.

So the danger of the blank spot and the opportunity of the blank spot come, in fact, from the same one thing: that it has no name yet. From the same root, the burden and the moat grow together. Then the next assignment is clear. What name to give this spot. What that name promises people and what it makes them misunderstand. Name it well and the burden crosses over into a moat; name it wrong and the moat slumps into a burden.

What This Chapter Did Not Answer

Chapter 3 is a record of finding the blank spot and tentatively concluding it is closer to opportunity than trap. But I did not attach the word tentative for nothing. Two things went unanswered.

One is what I said above. I have not yet crossed the river between wanting and opening a wallet. That the reason for the emptiness is gone is a supply-side story, and whether people will actually pay for uncomfortable learning is still an open question. This is not the kind of thing to conclude at a desk but the kind you only learn by making it and setting it in front of people. Until that day, this conclusion stays tentative.

The other is more concrete. In this chapter I said repeatedly that it fits nowhere, yet I have not properly shown what it differs from and how. The things that look most similar, say learning tools that lay AI over domains with right answers, companion apps that bond emotionally with people, light services that poke at curiosity. That mine does not go into one box with these I know by feel, but unless I take that feel apart point by point, I cannot answer the line “that is just such-and-such in the end.”

The next chapter does exactly that comparison. I will set the three that look closest side by side and pin down, one by one, the point where mine splits from them. Before naming the blank spot, it is the work of checking, against the neighboring boxes, whether the spot is really empty. If what I believed was a blank spot turns out to be a spot the next box already half occupies, one axis of this book topples whole. The next chapter is that check.